Software embebido

Using the process devised for Gradient Flow proved much more difficult and less informa-tive than K-mean and IsoData. The first issue was that the tile length had to be reduced to 300 pixels from the 400 pixel length used in the K-means clustering. This was due to the computation intensity required by Gradient Flow (see Section 5.2.2). Using a tile length of 300 pixels caused the entire process to take approximately 8 hours. Even though in-creasing the tile length would reduce the overall number of tiles, the computation amount increases exponentially with added pixels because each pixel must calculate its distance to all other pixels in the image. While 8 hours is not long relative to other remote sensing algorithms (such as unmixing), it does make the process of determining useful parameter values laborious. A technique used to facilitate parameter adjustment was running all

CHAPTER 6. RESULTS 63

the tiles using a particular set of Gradient Flow parameters and saving the output. After completing this process, which is the bulk of the processing time, the merging threshold could be altered repeatedly to discover the value that produced the most seamless results.

This technique obviously hinges on the selection of Gradient Flow parameters that provide the merging process with enough clusters to differentiate material classes, but not so many that continuity between tiles cannot be maintained. This task is not trivial; the image shown in Figure 6.3 found anywhere from 6 to 26 clusters for a given tile from a single set of Gradient Flow parameters. Being able to quickly alter the merging threshold through a range of values to generate Figure 6.3 and other classification maps, it was found that the Gradient Flow parameters have a much greater impact on producing seamless results.

Because of Gradient Flow’s unique method of clustering (using pixel density peaks), the merging technique that used means and standard deviations (see Equation 5.2) did not create class maps as seamless and diverse as K-means. The top excerpt from Figure 6.3 demonstrates both of these flaws. The reasoning for this behavior is that Gradient Flow works solely on histogram peaks so that it can potentially find clusters that are extremely close together as long as they are densely populated. K-means inherently spaces clusters apart because it moves means toward different pixels based on distance. Increasing the number of smoothing iterations can mitigate this effect, however, like attempting to use IsoData to generate flight line class maps, an added level of merging takes place, making it difficult to control two sets of extremely different merging parameters.

Using the insight provided by the parameter case study explained in Section 5.2.2, k and m values were chosen per tile to generate an average of 15 classes that could later be merged using the similarity metric thresholding technique. Figure 6.3 shows the result as well as describes the values used to achieve it. The merging threshold was set high to merge adjacent pixels from different tiles into a common class. The middle excerpt from Figure 6.3 shows that the merging technique does work to some extent. Two tiles join right at the center of that excerpt. The overall number of classes are low, but as shown by the bottom excerpt and flight line, very general yet important classes were differentiated. The red class highlights water from the Pacific Ocean collected at the bottom of the flight line.

The blue and orange classes are closely related as shown by the abrupt transition from one into another at tile seams, while both obviously representing vegetation. Generally, the blue class appears to be over darker vegetation than the orange class. Shown with more detail in Figure 6.4, the green class highlights all forms of exposed soil and minerals. The vegetationless peaks of mountains are highlighted, while the middle excerpt from Figure 6.3 shows that a white/grey mineral is clustered into the same class. The threshold could be decreased to differentiate structures plainly seen in the bottom color excerpt of Figure 6.3, but this would result in other classes being separated along tiles even though they are the same material. The purple class appears to highlight deep shadows. A small amount of dark blue class is present in the bottom third of the flight line, and correlates highly with vegetation growing at the base of those mountains. Whether this vegetation is

CHAPTER 6. RESULTS 64

Figure 6.3: Sample classification map generated using the Gradient Flow clustering al-gorithm and tile merging process. A random color assigned to each class. Each tile of 300 pixel length was run using parameters k=20 and m=8 producing anywhere between 6 and 26 classes for a given tile. The merging threshold was set at 1.1 DN producing a total of 6 classes for the entire flight line. Three excerpts are magnified along with their corresponding color renderings for closer viewing.

CHAPTER 6. RESULTS 65

Figure 6.4: Another higher magnified excerpt from Figure 6.3 showing the relationship between the orange class, green class, and the RGB color image.

truly different from those highlighted in blue and orange can not be concluded from visual interpretation. However, because the merging threshold is set so high, there must be a significant difference in their spectra. The peak normalized spectra (normalization occurs in Gradient Flow as well) are plotted for comparison in Figure 6.5. The main difference appears as a greater ratio of visible light to infrared. There is a small difference in spectral signature appearing in the green at 550 nm, where the orange class has a small peak. Both classes have surprisingly high values in the blue region of light. This is potentially due to atmospheric scattering.

Overall, the archaeologists have shown more interest in the following section’s results (spectral unmixing). With the addition of the computation time and current flaws, Gradi-ent Flow appears to be an unlikely candidate for further use in clustering Gradi-entire Hyperion flight lines. Partly for this reason, Gradient flow parameters for entire flight lines have not been examined more thoroughly and a better suited merging technique has not been developed. Creating such merging technique would be a difficult task as well, because Gradient Flow relies on pixel densities, hence any reduction in the pixel population used to handle an entire flight line has a detrimental effect on the algorithm’s clustering ability.